A two-stage optimized temperature component-driven deep learning model for arch dam deformation prediction

Current temperature component characterization methods cannot fully capture the time-varying characteristics of multi-point temperature sequences in arch dams. Thus, optimizing temperature component characterization and establishing accurate deformation prediction model are critical for the long-term health of arch dams that are significantly affected by thermal loading. This study proposes a deep learning model with a two-stage temperature factors optimization strategy for arch dam deformation prediction. In the first stage, a cosine similarity-based zoning method is employed to divide the dam temperature field into distinct zones. Kernel principal component analysis (KPCA) is applied in each zone to retain the leading principal components, which are then summed to form trend temperature factors. In the second stage, principal components with low contribution are evaluated via Chatterjee’s correlation coefficient (CCC) with dam deformation. Principal components whose CCC values exceed an established threshold are designated as discrete temperature factors. For the construction of the predictive architecture, the TransformerBiLSTM-based HT KC T model is established by the integration of Transformer and bidirectional long short-term memory (BiLSTM). The Transformer captures long-term temporal dependencies through its global modeling strength, while the BiLSTM analyzes temporal dynamics using its efficient local feature extraction capability. Hyperparameters of the model are optimized by the grey wolf optimizer. Results from a case study demonstrate that the developed model achieves superior precision and robustness than single models. The proposed HT KC T model outperforms the conventional HT T T model that relies on measured air temperature. The interpretation of the rationale behind the method is given by SHAP values.